1. Avoids allocating the list per proxy that was pretty bad.
2. Short-circuits inside the parallel FindPairs calls if the other proxy is also in the movebuffer so we don't do the fixture check in AddPair later in the single-threaded part. We also do this on the grid version so it should make grids moving near each other much faster by earlying out a lot earlier.
3. I wanted to put PreventCollideEvent in here too to get the really expensive one in but >turnstiles
4. Broadphase expand now checks realistically what it needs to check by considering the entity velocity (could make the bounds tighter by only checking the direction we came from but would rather just catch more collisions for now).

Want to get a profile before marking it as ready.

Old:

New:

Looks like doing the PreventCollideMessage inside or outside of the hot loop doesn't make a difference if the event itself has no subscribers so I might just leave that out for now to stop content shooting itself.